The AI terms that
actually matter.
Clear, technical definitions of 146 key concepts — written by AI engineers, not marketers.
Blackwell Architecture
NVIDIA's Blackwell architecture is a transformative GPU platform designed for trillion-parameter AI models, featuring a dual-reticle design, 208 billion transistors, and a 2nd-generation Transformer Engine with native FP4 precision.
Read definitionBM25 (Best Matching 25)
BM25 is the probabilistic ranking function that scores how well a document matches a query using term frequency, inverse document frequency, and document-length normalization. It's the default lexical scoring algorithm behind Elasticsearch, Lucene, OpenSearch, and the sparse half of nearly every hybrid search stack.
Read definitionCanary-Qwen 2.5B
NVIDIA's Canary-Qwen 2.5B is a Speech-Augmented Language Model that holds the #1 spot on the Hugging Face Open ASR Leaderboard with a 5.63% average word error rate.
Read definitionCatastrophic Forgetting
Catastrophic forgetting is the tendency of a neural network to abruptly lose previously learned knowledge or capabilities when it is trained on new data, because the same shared weights that encoded the old task get overwritten by gradient updates for the new one. It is the central obstacle to continual learning and a real risk in every LLM fine-tuning or RLHF pass.
Read definitionChain of Draft (CoD)
Chain of Draft is a prompting technique that keeps step-by-step reasoning but caps each step to a few words, so the model still works through a problem while emitting a fraction of the tokens of Chain-of-Thought. On several reasoning benchmarks it holds accuracy close to CoT while using as little as 7.6% of the tokens, which is why it has spread as a cost and latency cut.
Read definitionChain-of-Thought Prompting
Chain-of-Thought (CoT) prompting is a technique that instructs a language model to externalise its reasoning as a sequence of intermediate steps before producing a final answer, dramatically improving accuracy on arithmetic, logic, and multi-step tasks.
Read definitionChunk Overlap
Chunk overlap is the practice of letting adjacent retrieval chunks share a window of tokens or characters at their boundary, so a sentence or fact split across a fixed-size cut still appears intact in at least one retrieved chunk. It's the standard, low-effort fix for boundary information loss in fixed-size and recursive text splitters.
Read definitionColBERT (Late Interaction Retrieval)
ColBERT is a retrieval architecture that encodes queries and documents independently into one vector per token, then scores a pair at search time with a cheap operator called MaxSim, keeping token-level precision without paying the cost of a cross-encoder, and forming the basis of the late-interaction retrieval family (ColBERTv2, PLAID, ColPali, ColQwen).
Read definitionComposable AI
Composable AI is an architectural paradigm that replaces monolithic, rigid AI systems with a modular ecosystem of interoperable, reusable components—such as specific models, agents, and data pipelines—that can be assembled and scaled independently.
Read definitionComputer-Use Agent
A computer-use agent is an AI system that operates a computer's graphical interface the way a person does: viewing screenshots and issuing clicks, keystrokes, and scrolls, instead of calling app-specific APIs. Anthropic's Claude, OpenAI's Computer-Using Agent (Operator, ChatGPT agent), and Google's Gemini 2.5 Computer Use model are the three leading implementations.
Read definitionContext Engineering
Context engineering is the practice of deliberately curating the full set of tokens a language model sees at inference time: system prompt, tool definitions, retrieved documents, memory, and conversation history, not just the user prompt. Gartner called it the breakout AI capability of 2026, and it is the discipline that most separates a reliable agent from a flaky one.
Read definitionContext Rot
Context rot is the tendency for an LLM's reliability on a task to degrade as the input context grows longer, even well before the model's advertised context limit, and even when every piece of information the model needs is technically present in that context.
Read definitionStay ahead of the curve
Weekly newsletter on agentic AI, LLMs, and what we're building at Superteams — straight to your inbox.
Ready to ship AI in production?
We deploy fractional AI teams that deliver production-grade systems in 30–90 days. No fluff, no obligation.